If our dataset has 13 classes in which some classes have 10 images, some have 50 images, 2 classes have 100 images and 1 class have 200 images.so is our dataset imbalanced? And can this imbalanced dataset lead to over fitting? How can overcome this?

I tried dropouts regularization and several architecture of cnn and several hyper parameters.is there be any technique to know the number of layers of cnn. How many number of layers should be in cnn architecture and what is the architecture of cnn for this type of dataset. Our dataset have two folders in which one has query folder which has 13 classes and 2nd folder has document folder which has 13 classes similar to query folder and these 13 classes have number of images as stated above. My training accuracy is 100% and testing accuracy is 57%.how to reduce overfitting